一种利用人物相似度的视频索引算法.pdfVIP

第6 期 电 子 学 报 Vol .32 No. 6 2004 年6 月 ACTA ELECTRONICA SINICA June 2004 一种利用人物相似度的视频索引算法 1 2 2 1 王 鹏 , 马宇飞, 张宏江, 杨士强 ( 11 清华大学计算机科学与技术系, 北京 100084; 21 微软 洲研究院, 北京 100080) : 本文面向视频分析和索引技术, 提出一种利用人物相似度进行视频索引的算法. 该算法应用 SVMs 概 率输出理论, 将底层特征空间距离映射为语义层人物相似度, 并提出一种新的非监督聚类算法, 修正部分误判, 最终实 现对视频节目中人物的自动聚类和索引. 实验结果表明, 该算法实用而高效, 是对现有视频索引算法的有效补充. : 视频分析和索引; 人物相似度; 机器学习; 非监督聚类 : TP37 : A : 037222112 ( 2004) 062 09682 05 A Novel Video Indexing Approach Base d on P eople2Similarity 1 2 2 1 WANG Peng ,MA Yu2 ei , ZHANG Hong2j iang ,YANG Shi2qiang (11Department of Computer Scienceand Technology, Tsinghua University, Beijing 100084, China; 21Microsoft Research sia, Beijing 100080, China) Ab stract : This paper presents a novel approach to video indexing. In this approach, we de ine a people2similarity measure ac2 cording to both clothing similarity and sp eaking voice similarity. Such similarity depicts how perceptually similar two people app earing in di erent scenes are and whether they belong to an identical person .The extended support vector machines are applied to map a seri2 al o low2level eature distances to a perceived people2 similarity. To build people2 similar ity based video indexing, a novel unsupervised clustering algorithm is also proposed, which can more correctly group individual person according to the mutual people2 similarities a2 mong multip